Patch-Based Principal Component Analysis for Face Recognition

Area of Science:

• Computer Vision
• Machine Learning
• Biometrics

Background:

• Traditional Principal Component Analysis (PCA) methods for face recognition often overlook local spatial information.
• Pixel, column, or row correlations are commonly used, but may not capture essential facial features effectively.

Purpose of the Study:

• To propose a novel patch-based PCA method for enhanced face recognition.
• To leverage local spatial information within facial images for improved feature extraction.

Main Methods:

• Face images are divided into meaningful patches, treated as basic units.
• Patches are converted to column vectors, forming a new 'image matrix' for PCA.
• The two-dimensional PCA framework is adapted to compute correlations between patches, optimizing total scatter for feature extraction.

Main Results:

• The patch-based PCA method demonstrated improved accuracy on the ORL and FERET face databases.
• Experimental results show superior performance compared to one-dimensional PCA, two-dimensional PCA, and two-directional two-dimensional PCA.

Conclusions:

• Patch-based PCA is an effective approach for face recognition, outperforming existing PCA variants.
• Utilizing patches as fundamental units enhances the capture of local spatial information crucial for accurate face identification.